#Initialize pyprof
pyprof.init()
class Foo(torch.autograd.Function):
@staticmethod
def forward(ctx, in1, in2):
out = in1 + in2 #This could be a custom C/C++ function.
return out
@staticmethod
def backward(ctx, grad):
in1_grad = grad #This could be a custom C/C++ function.
in2_grad = grad #This could be a custom C/C++ function.
return in1_grad, in2_grad
#Hook the forward and backward functions to pyprof
pyprof.wrap(Foo, 'forward')
pyprof.wrap(Foo, 'backward')
foo = Foo.apply
x = torch.ones(4, 4).cuda()
y = torch.ones(4, 4).cuda()
with torch.autograd.profiler.emit_nvtx():
profiler.start()
z = foo(x, y)
profiler.stop()
def __init__(self, size):
super(Foo, self).__init__()
self.n = torch.nn.Parameter(torch.ones(size))
self.m = torch.nn.Parameter(torch.ones(size))
def forward(self, input):
return self.n * input + self.m
foo = Foo(4)
foo.cuda()
x = torch.ones(4).cuda()
#JIT the class using tracing
traced_foo = torch.jit.trace(foo, x)
#Initialize pyprof after the JIT step
pyprof.init()
#Assign a name to the object "traced_foo"
traced_foo.__dict__['__name__'] = "foo"
#Hook up the forward function to pyprof
pyprof.wrap(traced_foo, 'forward')
with torch.autograd.profiler.emit_nvtx():
profiler.start()
z = traced_foo(x)
profiler.stop()
print(z)
# # http://www.apache.org/licenses/LICENSE-2.0 # # Unless required by applicable law or agreed to in writing, software # distributed under the License is distributed on an "AS IS" BASIS, # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. # See the License for the specific language governing permissions and # limitations under the License. import torch import fused_layer_norm_cuda from apex.normalization import FusedLayerNorm import pyprof pyprof.init() pyprof.wrap(fused_layer_norm_cuda, 'forward') pyprof.wrap(fused_layer_norm_cuda, 'backward') pyprof.wrap(fused_layer_norm_cuda, 'forward_affine') pyprof.wrap(fused_layer_norm_cuda, 'backward_affine') input = torch.randn(20, 5, 10, 10).cuda() # With Learnable Parameters m = FusedLayerNorm(input.size()[1:]).cuda() output = m(input) # Without Learnable Parameters m = FusedLayerNorm(input.size()[1:], elementwise_affine=False).cuda() output = m(input) # Normalize over last two dimensions
import torch
from apex.optimizers import FusedAdam
import amp_C
import pyprof
pyprof.init()
# Wrap the custom fused multi tensor Adam implementation
pyprof.wrap(amp_C, 'multi_tensor_adam')
inp = 1024
hid = 2048
out = 4096
batch = 128
# Model
model = torch.nn.Sequential(
torch.nn.Linear(inp, hid).cuda().half(), torch.nn.ReLU(),
torch.nn.Linear(hid, out).cuda().half())
# Loss
criterion = torch.nn.CrossEntropyLoss().cuda()
# Adam optimizer
optimizer = FusedAdam(model.parameters())
# Input
x = torch.ones(batch, inp).cuda().half()
# Target
target = torch.empty(batch, dtype=torch.long).random_(out).cuda()
with torch.autograd.profiler.emit_nvtx():
y = model(x)
loss = criterion(y, target)
optimizer.zero_grad()
import torch
import torch.cuda.profiler as profiler
import pyprof
#The following creates an object "foo" of type ScriptModule
#The new object has a function called "forward"
@torch.jit.script
def foo(x, y):
return torch.sigmoid(x) + y
#Initialize pyprof after the JIT step
pyprof.init()
#Assign a name to the object "foo"
foo.__name__ = "foo"
#Hook up the forward function to pyprof
pyprof.wrap(foo, 'forward')
x = torch.zeros(4, 4).cuda()
y = torch.ones(4, 4).cuda()
with torch.autograd.profiler.emit_nvtx():
profiler.start()
z = foo(x, y)
profiler.stop()
print(z)
# You may obtain a copy of the License at # # http://www.apache.org/licenses/LICENSE-2.0 # # Unless required by applicable law or agreed to in writing, software # distributed under the License is distributed on an "AS IS" BASIS, # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. # See the License for the specific language governing permissions and # limitations under the License. import torch import fused_adam_cuda from apex.optimizers import FusedAdam, FP16_Optimizer import pyprof pyprof.init() pyprof.wrap(fused_adam_cuda, 'adam') model = torch.nn.Linear(10, 20).cuda().half() criterion = torch.nn.CrossEntropyLoss().cuda() optimizer = FusedAdam(model.parameters()) optimizer = FP16_Optimizer(optimizer) x = torch.ones(32, 10).cuda().half() target = torch.empty(32, dtype=torch.long).random_(20).cuda() y = model(x) loss = criterion(y, target) optimizer.zero_grad() loss.backward() optimizer.step()